Received July 8, 2019, accepted July 23, 2019, date of publication July 30, 2019, date of current version

August 14, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2931956

A Linear Model Based on Principal Component

Analysis for Disease Prediction
H. ROOPA 1 AND T. ASHA2
1 Department of Information Science and Engineering, Bangalore Institute of Technology, Bengaluru 560004, India
2 Department of Computer Science and Engineering, Bangalore Institute of Technology, Bengaluru 560004, India

Corresponding author: H. Roopa (roopatejas@gmail.com)

ABSTRACT Various classification methods are applied to predict different diseases, such as diabetes,
tuberculosis, and so on, in medical field. Diagnosis of diabetes can be analyzed by checking the level of
blood sugar of patient with the normal known levels, blood pressure, BMI, skin thickness, and so on. Several
classification methods have been implemented on diabetes. In this paper, the main aim is to build a statistical
model for diabetes data to get better classification accuracy. Extracted features of diabetes data are projected
to a new space using principal component analysis, then, it is modeled by applying linear regression method
on these newly formed attributes. The accuracy obtained by this method is 82.1% for predicting diabetes
which has reformed over other existing classification methods.

INDEX TERMS Principal component analysis, linear regression model, diabetes.

I. INTRODUCTION
Analysis of diseases is a tough task in medical field.
Diagnosing diabetes [2] can be understood by checking
the blood sugar level with normal desired level. In this
manuscript we present a statistical diabetes disease prediction
model where Principal Component Analysis (PCA) is applied
to extract attributes of Pima Indian Diabetes Data (PIDD) to
a new feature space. These attributes are then modeled using
linear regression model [8] to predict diabetes.
Attributes of PIDD are inspected at different angles to
obtain required information for processing data. So, feature
extraction is a major step in examining PIDD. The work
concentrates on retrieving feature values from PIDD to a
new feature space by employing PCA method. These new
set of feature values are inspected for their importance and
relevance, and are subjected for data mining methods like
LRM to classify the given data for predicting diabetes disease.
PCA is reduction method which considers the PIDD as
set of rows representing characteristics in a high dimensional
space and all rows are put up to a directions which represents
the best set of features. Here for original set of attributes
of PIDD, this transformation is applied to obtain an axis
that contains principle eigenvector where all the points of
all observations of each feature are spread out. Maximized
variance of data can be found on this axis. When considering
The associate editor coordinating the review of this manuscript and
approving it for publication was Yue Zhang.
second eigenvector, observe that axis along which the vari-
ance of distance from first axis is greatest and so on. Small
number of eigenvectors is represented by matrix of points,
then to minimize the root mean square error, approximation
of data is performed for the given number of columns in
the matrix consider. Thus the original features of PIDD are
approximated with fewer dimensions which are an overall of
original PIDD.
Model building provides a good fit to any set of data. Linear
statistical model estimate the unknown dependent PIDD fea-
ture value from the known independent PIDD feature values.
The representation of relationship between dependent PIDD
image feature and set of independent multiple PIDD features
are known as regression analysis.
The work is explained as follows, section 2 describes about
previous work on PIDD, proposed methodology is explained
in section 3, section 4 discusses about findings and section
5 gives about work’s conclusion.
II. RELATED WORK
Polat et al. [1] proposed a Least Square Support Vector
Machine (LS-SVM) classification method to obtain an accu-
racy of 79.16% which improvised over previous classifica-
tion methods. Generalized Discriminant Analysis (GDA) was
used at preprocessing stage for discriminating variables of
PIDD and then LS-SVM technique was applied on these
variables for classifying the disease.

105314 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/ VOLUME 7, 2019
H. Roopa, T. Asha: Linear Model Based on PCA for Disease Prediction
FIGURE 1. Proposed steps involved in feature extraction and modeling of
PIDD.
Seera and Lim [3] proposed Fuzzy Min–Max neural
network, Regression Tree and Random Forest (FMM-CART-
RF) combined hybrid classification method that achieved an
accuracy of 78.39% for PIDD.
Sa’di et al. [4] classified PIDD using various data mining
algorithms and analyzed that Naïve Bayes performed well
than RBF network and J48 with an accuracy of 76.95%.
Bansal et al. [5] proposed an evolutionary method where
feature selection of PIDD is obtained by Particle Swarm Opti-
mization (PSO) method and then k-Nearest Neighbor (KNN)
classification technique is applied on these features to achieve
an accuracy of 77%.
Choubey et al. [6] applied Genetic Algorithm (GA)
for variable selection on PIDD. Then implemented Naive
Bayes (NB) on these selected variables for classifying the
diabetes disease to obtain an accuracy of 78.69%.
III. PROPOSED METHODOLOGY
The proposed method includes extraction of new group of
features from PIDD by employing PCA so that the values
are inspected for their importance and relevance, and are
subjected for data mining methods like Linear Regression
Model (LRM) to classify the given data for predicting dia-
betes disease. The model is illustrated as shown in Figure 1.
A. INPUT DATA
The PIDD has been taken from UCI of machine learning
repository [9]. In this dataset female patient from Phoenix
and Arizona aged about 21 years are considered. The PIDD
consists of ‘8’ actual variables and ‘1’class variable. There
are totally 768 instances out of which 268 instances have class
variable value ‘1’ and 500 instances have class variable value
‘0’ respectively. The patients are classified as diabetes or
normal using binary valued variables. If the binary response
variable represents ‘1’ means ‘‘positive for diabetes’’ and ‘0’
means ‘‘negative for diabetes’’. The attributes of PIDD are
presented in Table 1.
VOLUME 7, 2019
TABLE 1. The eight attributes of PIDD with the class variable.
B. FEATURE EXTRACTION [7] OF PIDD BY APPLYING PCA
For the given numeric data matrix containing the feature
values of PIDD apply principal component analysis to project
the features to a new space.
The steps involved in feature extraction are
1) From each dimensions of PIDD subtract the mean,
which produces a data set whose mean is zero.
2) Calculate the variance matrix between two separate
dimensions of PIDD.
3) Find the Eigen vector and Eigen values of the matrix
obtained in step2.
4) Construct the feature vector and take transpose of it.
Then multiply it with original PIDD to obtain new set
of features projected to new space.
C. LINEAR REGRESSION MODEL (LRM)
Regression analysis predict dependent variable value ’y’
based on the set of independent variable values x1 , x2 , x3 ,. . . ,
xk.
The multiple independent variables analysis, linear
regression model [10] is represented by equation (1)
y = b0 + b1 x1+ b2 x2 + . . . + bk xk + ε (1)
where ‘y’ is the class variable of PIDD data,
b0 is a ‘y’ intercept,
ε is an error term and
b1 , b2 , b3,........... , bk are coefficients of x1 , x2 ,
x3, . . . . . . . . . .. . . ,xk. respectively.
The dependent variable ‘y’ of PIDD is the class variable,
which is predicted using equation (1). ‘y’ is based on the
105315
 H. Roopa, T. Asha: Linear Model Based on PCA for Disease Prediction

set of independent variable values x1, x2 , x3, . . . . . . . . . .. . . ,xk

represented by values of comp 1, comp 2, comp 3, comp 4,
comp 5, comp 6, comp 7 and comp 8 respectively.
The fitted line of equation (1) is calculated by principle
of least square method where b0 , b1 , b2 , b3,........... , bk are
chosen in such a way that the Sum of Squares of Error (SSE)
is minimum.
Consider equation (1), For n = 768 observation,
Let
 
y1
 y2 
 
Y=  y3 

 : 
y738
FIGURE 2. LRM results of PIDD.
1 x11 x21 . . .
 
x81
X =  ... ..
.
.. 
. 


1 x1786 . . . x8768
 ˆ 
b0
 bˆ1 
 
 bˆ 
b̂ =  2
 : 
 
 : 
bˆk

b̂ → least square estimates of b0 , b1 , . . . , bk of linear

model.
Least square matrix is obtained by
(x1 x)b̂ = x1 y
where x1 → is transpose of x matrix.
(x1 x) → coefficient matrix of least square estimates
of bˆ0 , bˆ1 . . . . . . . . . bˆk .
x1 y → gives matrix of constants.
Therefore least square solution is obtained by b̂ = (x1 x) −1 .
1
x y
Substitute b̂ in equation (1) to get the final fitted line.
D. OUTPUT DATA
The model classifies features of PIDD as diabetic or normal.
IV. RESULTS
The attributes of PIDD is projected to a new space using PCA.
Then LRM is applied on the components of PIDD. The results
of the model are presented in Figure 2.
The residuals error between the prediction of LRM and
actual results is smaller ranging between −1.0059 and
1.2930. And the median value is closer to zero. F-statistics
shows that model has at least one variable that is significantly
different than zero. The variables comp1, comp2, comp5,
comp6, comp7 and comp8 are more significant indicated by
∗∗∗ and ∗∗ when compared to comp3 and comp4 respectively,
which are less significant.
The components of PIDD are partitioned, where 80% are
considered for training the model and 20% of data for test-
ing the model respectively. These data are classified using
FIGURE 3. Performance of training data.
LRM. The actual and predicted values of PIDD modeled by
LRM are given by confusion matrix. The confusion matrix
and statistics of training and testing data of PIDD are given
in Figure 3 and Figure 4 respectively.
The confusion matrix of training data as given in
Figure 3 depicts that 614 instances out of 768 instances
of PIDD are considered for training. 351 observations are
correctly predicted negative for diabetes and 110 observations
are correctly predicted positive for diabetes. So, 461 observa-
tions have been correctly classified by LRM. 46 observations
represent incorrect prediction of LRM that resulted positive
for normal person and 99 observations represent incorrect
prediction of LRM that resulted negative for diabetic person.
So, 145 observations have been wrongly classified by LRM.
The confusion matrix of testing data as given in
Figure 4 depicts that 154 out of 768 instances of PIDD
are considered for testing. 96 observations are correctly pre-
dicted negative for diabetes and 37 observations are correctly
predicted positive for diabetes. So, 133 observations have
been correctly classified by LRM. 7 observations represent
incorrect prediction of LRM that resulted positive for normal
person and 22 observations represent incorrect prediction
of LRM that resulted negative for diabetic person. So, 29
observations have been wrongly classified by LRM.

105316 VOLUME 7, 2019

H. Roopa, T. Asha: Linear Model Based on PCA for Disease Prediction
FIGURE 4. Performance of testing data.
TABLE 2. The performance evaluation of training and testing data using
PCA-LRM on PIDD.
The performance of LRM can be analyzed by its accuracy
in prediction of disease. The implementation of LRM on
PIDD uses the concept of probability to reveal prediction of
diabetic or nondiabetic patient. The default probability value
considered is 0.5 which is the cutoff value. If the probability
value falls below 0.5 then the possible LRM prediction is neg-
ative for diabetes otherwise if it is above 0.5, LRM predicts
the patient to be diabetic.
The overall performance of the model on PIDD is
represented in Table 2.
From PIDD, 80% of data is considered for training the
data. The snapshot of performance curve of training data of
PIDD is shown in Figure 5. The performance curve of LRM
is analyzed with respect to accuracy and cutoff value used in
disease prediction. Figure 5 depicts that at cutoff value 0.5,
LRM prediction accuracy raises above 70%. LRM prediction
accuracy falls low when the cutoff value is below 0.5.
The Receiver Operating Characteristics (ROC) curve of
LRM demonstrates ability of classification based on the vari-
ation of cutoff values. The construction of ROC curve at
various cutoff values is plotted by True Positive Rate (TPR)
VOLUME 7, 2019
FIGURE 5. Training data’s Performance curve.
FIGURE 6. Training data’s ROC curve.
FIGURE 7. Testing data’s Performance curve.
against False Positive Rate (FPR). Here TPR is named as
Sensitivity and calculated FPR is known as (1- Specificity).
Hence ROC curve is used to analyze the optimality of LRM.
The ROC curve is interpreted by calculating value of Area
Under Curve (AUC). The overall ROC curve value is 1 which
is the cumulative probability distribution value. A slope is
drawn and used as a partition to analyze a ROC curve. The
ROC curve if present above the slope indicates that LRM
prediction is more than 50% where as if ROC curve if present
below the slope then LRM prediction is less than 50% which
is not good for a model.
The snapshot of ROC curve of training data of PIDD is
shown in Figure 6. In this Figure 6 the prediction ROC curve
is distributed over various cutoff values ranging between
−0.53 and 0.82. The LRM prediction of ROC curve is more
than 50% as the curve is above the slope. It is observed from
105317

FIGURE 8. Testing data’s ROC curve.
TABLE 3. Comparison of outcomes of PIDD with several existing
methods.
the Figure 6 that near 0.5 cutoff value, TPR value is high. The
calculated AUC value of training data is 0.8196.
From PIDD, 20% of data is considered for testing the data.
The snapshot of performance curve of testing data of PIDD
is shown in Figure 7. The performance curve of LRM is
analyzed with respect to accuracy and cutoff value used in
disease prediction. Figure 7 depicts that at cutoff value 0.5,
LRM prediction accuracy raises above 80%. LRM prediction
accuracy falls low when the cutoff value is below 0.5.
The snapshot of ROC curve of testing data of PIDD is
shown in Figure 8. In this Figure 8 the prediction ROC curve
is distributed over various cutoff values ranging between
105318
H. Roopa, T. Asha: Linear Model Based on PCA for Disease Prediction
−0.19 and 1.01. The LRM prediction of ROC curve is more
than 50% as the curve is above the slope. It is observed from
the Figure 8 that near 0.5 cutoff value, TPR value is high. The
calculated AUC value of testing data is 0.8487.
There are various existing, implemented methods on
PIDD. Few are mentioned in Table 3, which gives details
about the performance measure of all the executed methods.
It is analyzed from Table 3 that PCA-LRM achieves improved
accuracy on PIDD than other prevailing methods.
V. CONCLUSION AND FUTURE WORK
Feature extraction and statistical modeling on PIDD is
presented in this research work. The PIDD features are
extracted to a new space using PCA. These newly projected
features are then modeled using LRM to predict whether the
patient is diabetic or normal. The results obtained in this
study have achieved high accuracy rate for predicting diabetes
when compared with other existing methods. The proposed
statistical model can be adopted for predicting different kinds
of diseases like tuberculosis, eye disease, cancers, etc.,
REFERENCES
[1] K. Polat, S. Güneş, and A. Arslan, ‘‘A cascade learning system for
classification of diabetes disease: Generalized discriminant analysis and
least square support vector machine,’’ Expert Syst. Appl., vol. 34, no. 1,
pp. 482–487, 2008.
[2] M. F. Ganji and M. S. Abadeh, ‘‘A fuzzy classification system based on
ant colony optimization for diabetes disease diagnosis,’’ Expert Syst. Appl.,
vol. 38, no. 12, pp. 14650–14659, 2011.
[3] M. Seera and C. P. Lim, ‘‘A hybrid intelligent system for medical data
classification,’’ Expert Syst. Appl., vol. 41, no. 5, pp. 2239–2249, 2014.
[4] S. Sa’di, A. Maleki, R. Hashemi, Z. Panbechi, and K. Chalabi, ‘‘Compari-
son of data mining algorithms in the diagnosis of type II diabetes,’’ Int. J.
Comput. Sci. Appl., vol. 5, no. 5, pp. 1–12, 2015.
[5] R. Bansal, S. Kumar, and A. Mahajan, ‘‘Diagnosis of diabetes mellitus
using PSO and KNN classifier,’’ in Proc. Int. Conf. Comput. Commun.
Technol. Smart Nation (IC3TSN), Oct. 2017, pp. 32–38.
[6] D. K. Choubey, S. Paul, S. Kumar, and S. Kumar, ‘‘Classification of Pima
indian diabetes dataset using naive Bayes with genetic algorithm as an
attribute selection,’’ in Proc. Int. Conf. Commun. Comput. Syst. (ICCCS),
Feb. 2017, pp. 451–455.
[7] H. Roopa and T. Asha, ‘‘Feature extraction of chest X-ray images and
analysis using PCA and kPCA,’’ Int. J. Elect. Comput. Eng., vol. 8, no. 5,
p. 3392, Oct. 2018.
[8] H. Roopa and T. Asha, ‘‘Analysis of feature ranking methods on
X-ray images,’’ in Proc. Int. Conf. ISMAC Comput. Vis. Bio-Eng. Cham,
Switzerland: Springer, 2018, pp. 1393–1403.
[9] UCI Repository of Machine Learning. Accessed: 2017. [Online]. Avail-
able: http://www.ics.uci.edu./~mlearn/ML Repository.html
[10] H. Roopa and T. Asha, ‘‘An efficient prediction model to analyze tubercu-
losis chest X-ray images,’’ Int. J. Sci. Eng. Res., vol. 10, no. 1, pp. 694–700,
2019.
[11] D. Sisodia and D. S. Sisodia, ‘‘Prediction of diabetes using classification
algorithms,’’ Procedia Comput. Sci., vol. 132, pp. 1578–1585, Jan. 2018.
[12] A. Ahmad, A. Mustapha, E. D. Zahadi, N. Masah, and N. Y. Yahaya,
‘‘Comparison between neural networks against decision tree in improving
prediction accuracy for diabetes mellitus,’’ in Proc. Int. Conf. Digit. Inf.
Process. Commun. Berlin, Germany: Springer, 2011, pp. 537–545.
VOLUME 7, 2019